In the AI era, there are new experiences every day. In this age of technological explosion, we face disruption and reconstruction daily.
Recently, clawdbot has gone viral in Silicon Valley, even becoming a phenomenon that sold out Mac minis as an open-source AI agent.
What makes it different?
<h2>1. What exactly is clawdbot?</h2>
In one sentence: Clawdbot = &#34;self-hosted Claude assistant + local automation operating system.&#34; It is not a website but an AI agent framework running on your own machine.
Several typical descriptions from the official and community:
&#34;A personal AI assistant that runs on your own device,&#34; communicating with you through various IM platforms like WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, WebChat, etc.
The core idea is: the Gateway is just the control plane; the real product is the &#34;assistant that can actually get things done&#34; itself—able to clear your inbox, send emails, manage your calendar, check flights, and help you check in.
Many articles directly refer to it as &#34;Claude with hands&#34;: it can chat, run scripts, control browsers, and manipulate the file system.
<h2>2. Who developed it? What is the background?</h2>
Author: Peter Steinberger (@steipete)
An Austrian software engineer, previously best known as the founder of PSPDFKit—a PDF SDK for iOS/Android/Web. The company later rebranded to Nutrient and was acquired by private equity, making him a &#34;financially free retired dev entrepreneur.&#34;
Before Clawdbot emerged, he wrote a popular long article titled &#34;Claude Code is my computer,&#34; discussing how he used Claude Code as a &#34;remote operating system&#34; to let the model control an entire macOS. Clawdbot is essentially the product of systematizing and open-sourcing that personal experiment.
The timeline is roughly as follows:
<ul>
<li>End of 2025: He first created a personal &#34;Clawd assistant&#34; using Claude, a space lobster (the project's mascot).</li>
<li>January 2026: Officially open-sourced the project under the name Clawdbot, launching on GitHub and the clawd.bot website; within weeks, GitHub stars increased from a few thousand to twenty to thirty thousand, recently surpassing forty thousand stars, becoming one of the fastest-growing open-source projects currently.</li>
</ul>
In a short time, it attracted dozens of contributors and a Discord community of thousands; many security experts and AI influencers (including Karpathy, etc.) have discussed it on X
In simple terms: This is a project about &#34;a retired PDF SDK expert returning to create an AI agent,&#34; not an anonymous small toy.
<h2>3. Why has Clawdbot suddenly become popular? (A few key points behind it)</h2>
Operating mode: Fully self-hosted, local first
Clawdbot is open-source under the MIT license, and you can run it on your own devices: Mac, Linux, Windows (including WSL2), Raspberry Pi, or even VPS.
Default recommended scenarios:
A M-series Mac mini at home (low power consumption running the agent 24/7);
Or a $5 per month VPS in the cloud (like DigitalOcean/Hetzner).
The model can connect to:
Anthropic Claude (officially recommended Pro/Max + Opus 4.5); OpenAI, MiniMax, etc.; and can also connect to local LLMs through Ollama and others.
Result: Data and automation logic are all &#34;on your side,&#34; not on a SaaS platform; the cost structure becomes &#34;hardware + API fees&#34; (Claude / OpenAI), while Clawdbot itself is free; this aligns very well with the sentiment of &#34;developers holding the infrastructure themselves&#34; and &#34;privacy control.&#34;
Capability boundaries: From &#34;chatting&#34; to &#34;doing work&#34;
Compared to traditional chatbots, it has a few key points:
① Unified multi-channel entry
The same assistant can be simultaneously connected to WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Microsoft Teams, Google Chat, WebChat, etc.; for you, it's &#34;one person,&#34; but technically it's a multi-channel inbox + session model.
② Continuous memory &amp; proactive awakening
Clawdbot has long-term memory (Markdown/vector retrieval, etc.), and won't forget like ChatGPT when you close the tab; it can send you morning briefings, urgent reminders, monitor emails/market conditions, and proactively ping you.
③ Real &#34;hands-on ability&#34;
With tools + shell permissions, it can execute system commands, write scripts, manipulate files, control browsers, help you log into websites, send emails, modify To-Do lists, control smart home devices, etc.; this is also why the title of that article in the security circle is &#34;Clawdbot is what happens when AI gets root access.&#34;
④ Multi-agent orchestration
Supports running multiple dedicated agents: coding agent, research agent, life affairs agent, collaborating through routing; internally has a multi-agent routing, subagent, and tool sandbox architecture
<h2>4. Tech Stack &amp; Architecture</h2>
Clawdbot is essentially a complete agent runtime case:
The underlying infrastructure (Mac / VPS) supports LLM providers (such as Anthropic / OpenAI), connects to various APIs/websites externally, and has memory, skills, and hooks internally.
Tech Stack Overview: A &#34;structural diagram&#34; is established first.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ac3c6e58-e019-4942-aff5-8588f3e437d3.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ac3c6e58-e019-4942-aff5-8588f3e437d3.png"/>
Using the familiar tech stack diagram with L0–Lx layers, it is abstracted as:
L0: Compute &amp; Network
Mac mini / VPS / home server + public network / VPN / Tailscale. L0 is a &#34;light CPU / high connection count / I/O intensive&#34; form, rather than a GPU node.
L1: OS &amp; Runtime
macOS / Linux / WSL2 + Node.js / Bun + systemd / daemon. L1 determines reliability and recoverability: using systemd / launchd / daemon mode to ensure automatic restart of the gateway after a crash; the local file system is used for persistent config / memory / logs.
L2: Gateway &amp; Control Plane (Core of Clawdbot)
CLI, Gateway process, configuration system, queue, health, logging. The Control Plane / Service Mesh of the Agentic OS provides routing, rate limiting, health checks, configuration, and discovery for the upper layer's agent loop / tools / channels.
L3: Agent Loop &amp; Context Layer
Agent definition / system prompt / sessions, queues, multi-agent routing, tool policy. L3 is the &#34;Agent Behavior Model Layer,&#34; which inherently supports multi-agent collaboration, tool use, and long-running loops, making it closer to real production workflows than traditional &#34;LangChain-style single-round tool calls.&#34;
L4: LLM &amp; Provider Abstraction Layer
Anthropic / OpenAI / Bedrock / OpenRouter / local models (such as Ollama), including failover and usage tracking. This layer is key for model routing / failover / cost optimization: the same agent can switch different models for different tasks (for example, Claude for reasoning)
Cheap models handle bulk tasks), even running behind middleware like Vercel AI Gateway.
L5: Tools / Skills / MCP / Browser / Exec Layer
HTTP tools, browser, file system, plugins, Clawdhub skills, MCP / external tools, in the TechStack model, this layer is: Workload-Specific Tools / Skill Fabric—determining the extent to which an agent can truly &#34;get things done&#34; rather than just chat.
L6: Channels &amp; Surfaces
WhatsApp / Telegram / Slack / Discord / Signal / iMessage / WebChat / TUI / Voice, L6 determines the &#34;entry form&#34; of the agent: it can wait for user input like a traditional ChatBot; it can also proactively send messages using channels &amp; hooks (an agent that &#34;will reach out to you first&#34;).
L7: Automation &amp; Hooks &amp; Memory
Cron, Webhook, Gmail Pub/Sub, Auth monitoring, Markdown memory, compaction, session pruning, L7 is Stateful Automation + Memory Fabric—determining whether the agent can evolve from a &#34;one-time tool&#34; to a &#34;long-term personal assistant / background process.&#34;
There are three cross-cutting concerns:
Security (sandbox vs elevated, OAuth, API keys, hardening),
Observability &amp; Ops (doctor, status, logs, health, heartbeat),
Remote / multi-gateway / discovery (Tailscale, remote gateway, Bonjour).
Three key points from &#34;Tech Stack&#34; to &#34;agentic ecological niche&#34;
Breaking down this stack, it can be more precisely stated:
① Clawdbot itself is an &#34;Agent Runtime × Control Plane&#34; ecological niche:
L2–L4 unifies &#34;models + tools + channels + memory&#34; within a single runtime; technically it can take on any LLM provider and tools, positioned similarly to a &#34;local version of LangGraph × multi-channel gateway × MCP client.&#34;
② Around the &#34;traffic infrastructure&#34; niche of Clawdbot / agentic:
LLM becomes the web gateway (Cloudflare AI Gateway, Vercel AI Gateway, etc.); WAF / Bot Management / AI Security Suite on the API / website side is used to block these agentic traffic and prompt injection. The token / API explosion brought by the L4 model layer &amp; L5 tools layer will naturally give rise to:
③ Security &amp; hardening niche:
The community has already pointed out in security reports and Reddit discussions that improper configuration can directly dump sensitive information such as Anthropic API keys; this means that in addition to Clawdbot's own security documentation, there will also be a dedicated hardening toolchain, policy engine, Zero Trust for AI, and a complete stack for enterprises.
What are the mapped investment opportunities?
<ol>
<li>Consumer / Personal: End-side AI + Agent Host Rail</li>
</ol>
Clawdbot + Mac mini prove that high-frequency, private, personalized Agent work can run continuously on the end side (Mac / AI PC / high-end mobile phone). Cloud training and global large models continue in super-large data centers; however, inference + personalization + system-level functions are increasingly being pushed to the end-side NPU/unified memory.
(1) End-side complete machine / SoC control players
Apple: Controls M/A series SoC + unified memory + macOS/iOS + Apple Intelligence, directly tying &#34;the ability to use AI functions&#34; to hardware thresholds → upgrade barriers.
Windows Copilot+ PC ecosystem: MSFT + Qualcomm/Intel/AMD, &#34;No NPU 45+ TOPS is not a Copilot+ PC,&#34; marking the standardization of NPU in PCs.
<blockquote>
Key: Whoever can make AI functions into &#34;hardware gating + OS deep integration&#34; will have a higher ASP + stronger control over upgrade cycles.
</blockquote>
General Edge AI hardware / MCU supply chain
Overall Edge AI hardware: From approximately $26.1 billion in 2025–2030 to approximately $58.9–68.8 billion, CAGR ≈ 17–18%. The IoT MCU market is expected to be approximately $7.3 billion by 2030, with a CAGR of approximately 6.3%, driven in part by industrial/consumer Edge AI
The opportunities here are more dispersed, suitable to serve as a &#34;client-side AI hardware Beta pool&#34; rather than a single all-in:
Mobile/PC SoC (Qualcomm, Apple, Intel/AMD NPU);
Various Edge SoC / MCU / ISP (NXP, TI, Ambarella, etc.).
<ol>
<li>Cloud / Network: AI Gateway &amp; Efficiency Layer Rail</li>
</ol>
In the world of agents, &#34;computing power is expensive + invocation is dense&#34; → there must be an AI Gateway + edge inference platform: unified routing, caching, rate limiting, cost measurement; filtering invalid/malicious requests; making a judgment at the edge/PoP before deciding whether to hit the most expensive large model.
Representative players &amp; opportunities
(1) Cloudflare (NET) type: AI toll station + efficiency layer
AI Gateway: consolidates all LLM calls, performs logging, routing, rate limiting, caching, cost modeling.
Workers AI / Agents: run inference directly at the Cloudflare edge, charging based on &#34;neurons&#34; / requests.
<blockquote>
Alpha is not just an additional piece of HBM, but rather: under the premise of given GPU/power, using routing/caching/multi-model selection/security to turn unit computing power into more effective Agent work.
</blockquote>
(2) Tools providing &#34;AI observation &amp; budget control&#34;
Datadog LLM Observability + Agent Monitoring:
Trace each agent link: input/output/latency/token/errors/tool calls/inference steps; helping enterprises see &#34;which uses the cloud and which can be pushed down to the edge/client.&#34;
<blockquote>
NET / DDOG can be viewed as the same &#34;efficiency + governance rail&#34;:
One stands at the network edge, and the other at the observability/SecOps middle platform,
both benefiting from the long-cycle increment of Agent/LLM usage rather than directly betting on GPU quantity.
</blockquote>
<ol>
<li>AI Security &amp; Identity: Agent Attack Surface Rail</li>
</ol>
AI agents = &#34;machine identities&#34; with keys, limbs, and memory;
Prompt injection/data leakage/tool abuse/session smuggling become new attack surfaces;
Enterprises must open new budget pools for identity + Zero Trust + runtime security + AI layer protection.
<ol>
<li>Representative players &amp; opportunity mapping</li>
</ol>
(1) Runtime &amp; LLM/Agent security
Palo Alto Networks (PANW) Prisma AIRS: AI security platform, clearly protects against prompt injection, data leakage, malicious code, model abuse, etc.
CrowdStrike (CRWD) Falcon AIDR: AI Detection &amp; Response, specifically looks at the prompt layer + Agent usage, providing visibility and automated response.
(2) Identity &amp; Zero Trust &amp; Channel Control
CRWD acquired SGNL, expanding &#34;continuous identity assessment&#34; to machine/AI identity, which is a necessary layer in the Agent era;
Zscaler / Okta / CyberArk, etc.: control &#34;which Agent / Service can access which APIs / databases / SaaS,&#34; acting as a gate in a multi-Agent/multi-tool topology.
<blockquote>
Investment Mapping:
Pull these companies out of the traditional &#34;Sec SaaS&#34; basket and view them as the core of the &#34;Agent Attack Surface Rail&#34;:
<ul>
<li>Directness: whose products directly mention &#34;LLM/Agent security&#34;;</li>
<li>TAM: who has a complete chain of identity + network + runtime.</li>
</ul>
</blockquote>
IV. Data &amp; Agent OS: Data Brain + Agent Platform Rail
For Agents to truly &#34;understand the business,&#34; they need clean, centralized, well-governed data;
Going a step further: the data platform must transform into an Agent orchestration platform, otherwise it is just providing blood supply to others.
<ol>
<li>Representative Players &amp; Opportunities</li>
</ol>
Snowflake (SNOW) Logic
Data Cloud = Enterprise Data Brain;
Cortex AI Functions + Cortex Agents + Snowflake Intelligence = Embedded Agent Framework;
Collaboration with Anthropic for $200 million over the years on &#34;Agentic AI,&#34; emphasizing high-regulation industries (finance/healthcare) using SNOW as the core for &#34;data + Agent OS.&#34;
<blockquote>
Investment Perspective:
<ul>
<li>SNOW is not just a &#34;data warehouse,&#34; but is competing for the &#34;Enterprise Agent OS&#34; stack position,</li>
<li>Directness very high: complete Agent Framework + Data integrated;</li>
<li>TAM = original data warehouse/analytics + AI/Agent platform.</li>
</ul>
</blockquote>
Other players of similar type will emerge (Databricks, some cloud-native data platforms), but SNOW currently tells the most aggressive &#34;Agent + enterprise data integration story.&#34;
V. Industry: AI Edge Controllers / Industrial Agent Rail
Industrial Control (PLC/DCS): Extremely high requirements for real-time performance, determinism, reliability, and compliance;
AI models must participate in the control loop and can only operate closely on-site: edge controllers;
The cloud is responsible for: training, cross-plant analysis, digital twins, unified management;
Edge controllers are responsible for: online inference for quality inspection, predictive maintenance, process optimization, anomaly detection, etc. + real-time closed loop.
<ol>
<li>Key Actions of Siemens / ROK, etc.</li>
</ol>
(1) Siemens
Industrial Edge + AI Inference Server: Edge app form deployment model inference, tightly coupled with PLC/vPLC. S7-1500V virtual PLC: control logic itself can run in vPLC form in the Edge environment, achieving a three-in-one of PLC + Edge server + gateway.
(2) Rockwell (ROK)
Embedded Edge Compute Module: Plugged into the Allen-Bradley control platform, providing local edge computing power and gateway capabilities. Collaborating with NVIDIA: Nemotron Nano SLM → FactoryTalk Design Studio / Copilot, running small language models on the industrial Edge to assist engineering/maintenance repairs and design.
Investment Mapping for Industrial Rail:
<ul>
<li>Directness: Factories that integrate control + Edge platform + industrial AI (Siemens / ROK / Schneider / ABB…);</li>
<li>TAM: Upgrading from &#34;only selling PLCs&#34; to &#34;selling AI-ready Edge controllers + Edge platform licenses + cloud services,&#34; significantly increasing the lifecycle value of single devices;</li>
<li>The growth of this rail will shift some AI inference from the cloud to on-site, while simultaneously increasing the demand for training, digital twins, and cross-plant optimization on the cloud side.</li>
</ul>
Clawdbot + Mac mini have shown that edge personal computing power can handle a large number of high-frequency, privacy-sensitive, personalized agent workloads, which will shift part of the &#34;CapEx of the AI era&#34; from data centers to terminals and edges.
However, in terms of scale and workload types, it will more change the structure of CapEx in the short to medium term (more efficiency layers and edge-side SoC investments), rather than causing the overall CapEx curve of hyperscalers to turn downward.
The real systemic risk will only appear in an extreme world: small models + local NPUs are sufficient to cover the vast majority of tasks, while regulations and costs force all industries to adopt a local-first approach, and cloud providers fail to timely transform into model licensing and edge orchestration platforms. The current realistic path resembles &#34;cloud-edge-end collaboration → AI total CapEx operating at a high level,&#34; rather than &#34;end-side consuming cloud.&#34;
Risk Warning and Disclaimer
The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk.

Clawdbot is an open-source AI agent that allows users to run a personal AI assistant locally, supporting various instant messaging tools. Its developer is Austrian software engineer Peter Steinberger, who founded PSPDFKit. Clawdbot combines Claude's chat capabilities with automation, enabling it to handle tasks such as emails and calendars, becoming a smart assistant for users

Clawdbot + Mac mini: When AI assistants run locally